This seems like a rather normal process yet there is no way (as a body of a parallel_for), that you can get a handle on the root task, start_for. So this means you can't allocate a child task.
A simple example (that has come up) is a parallel_loop that pushes on samples from an array that satisfies a function. If pushing on the sample takes a substantial amount of time, a mutex won't be ideal, but spawning/recycling a new task that will handle it would be best. Originally, I had designed it so that there was a task that just called a for loop - and pushed across a boundry, but this doesn't seem like the best solution.
More like a feature request - is this even possible now, reason why it won't be like this in the future?